Security risk report for @forgecat/anthropics_claude-plugins-official_math-olympiad v0.0.5
Source Integrity
Low
Profile is attributed to Anthropic and references the official claude-plugins-official repository with a specific commit hash (3deb821cb71ccfaaf2ffa9935e977df314ce5cd5).
All external references (GitHub, arXiv) point to legitimate, well-known sources; no typosquatted or attacker-controlled package names are present.
The profile declares no dependencies and no MCP servers, reducing supply-chain risk.
Agent Intent
Low
The profile describes a legitimate mathematical problem-solving and verification system with no instructions to manipulate the AI, exfiltrate data, or hide its own instructions.
Content is educational and methodological: it documents adversarial verification patterns, solver heuristics, and presentation techniques for competition mathematics—all standard pedagogical material.
The adversarial verifier prompts (e.g., 'Pattern #4: Would it prove too much?') are _descriptions_ of verification techniques, not instructions to the AI to perform malicious actions; they teach the AI how to critique proofs, not how to compromise itself.
README and SKILL.md describe the profile's purpose transparently: 'Solve competition math problems with adversarial verification.'
Adversarial prompts are framed as checks the verifier should run (e.g., 'Your goal is to find ANY reason to refute'), not as instructions to ignore system prompts or exfiltrate data.
No instruction to read credentials, transmit system prompts, install remote payloads, or weaken security defaults.
Permissions
Low
The profile declares no tools, no file access, no shell execution, and no MCP servers—only pure reasoning (thinking/language model output).
The stated function is mathematical problem-solving and proof verification, which requires only reasoning capability, not external agency.
Subagent invocations (solver and verifier agents) are scoped to mathematical reasoning tasks with explicit constraints ('NO COMPUTATION. Do not use Bash, Python, WebSearch, Read, Write, or any tool that runs code or fetches data').
Details
Evidence
SKILL.md: 'Uses pure reasoning (no tools)' and 'Tool policy: Solvers and verifiers use THINKING ONLY.'
attempt_agent.md: 'You have NO tools — pure reasoning only.'
No file_write, file_delete, shell, or web_fetch permissions requested or implied.
MCP Risk
Low
No MCP servers are defined in the profile; the skill operates entirely within the language model's reasoning capability.
No external binary execution, network access, or filesystem operations are configured.
The profile references external documentation (GitHub, arXiv) but does not invoke or configure any MCP server to fetch or execute code from those sources.
ForgeCat